Updates for PyTorch 1.2 masking/bool behavior

Summary: Pull Request resolved: fairinternal/fairseq-py#821

Differential Revision: D16790120

Pulled By: myleott

fbshipit-source-id: 2fb5070172636561d08596a29f08c93df07548bf

fairseq/data/mask_tokens_dataset.py

@@ -127,7 +127,7 @@ def __getitem__(self, index: int):
                 if self.mask_whole_words is not None:
                     mask = np.repeat(mask, word_lens)
                 new_item = np.full(len(mask), self.pad_idx)
-                new_item[mask] = item[torch.from_numpy(mask.astype(np.uint8))]
+                new_item[mask] = item[torch.from_numpy(mask.astype(np.uint8)) == 1]
                 return torch.from_numpy(new_item)
 
             # decide unmasking and random replacement